Abstract:
Although numerous open-source tools exist for machine learning with tabular data, there is a scarcity of comparable resources tailored specifically for NLP. The lack of transparency in the inner workings of existing AutoNLP tools is a significant obstacle in scientific research. AutoML tools are known for their ability to perform model selection with little human intervention, improving the accuracy and reliability of the results. However, conducting a model space search among pre-trained NLP models can be computationally infeasible, making it challenging to determine the optimal NLP model for a given dataset. This research aims to enhance the performance of NLP model selection. Our approach has resulted in higher accuracy than existing methods on the dataset that was created. In our future work, we plan to benchmark our algorithm against datasets created by other researchers to validate its effectiveness. Additionally, we intend to use the same system to perform model selection among popular large language Transformer models.
